Text-to-SQL-using-T5-Model

In this project, we have fine-tuned T5, a transformer-based large language model developed by Google, and pre-trained on a large set of varied text datasets including Wikipedia, news articles, books, etc, to work on a bunch of NLP tasks like Summariza- tion, Question Answering, Text completion, etc. We evaluated our method on the SPIDER dataset [15], a benchmark dataset for Text-to-SQL tasks. Fine-tuning T5 on Spider, we were able to get significant perfor- mance improvement of at least 16%, increasing from 80% to 93%. We, further, implemented rule-based ap- proaches as a post-processing step to rectify common errors produced by the model, which further improved our scores to 95%.

Final Report Link